class ReseauSimple:
def __init__(self, archi=[2,3,1], xmod=False):
if xmod:
self.X = np.random.randn(archi[0], 1)
self.W_1 = np.random.randn(archi[1], archi[0])
self.W_2 = np.random.randn(archi[1], archi[2])
self.b_1 = np.random.randn(archi[1], archi[2])
self.b_2 = rd.normalvariate(mu=0, sigma=1)
self.stock = [[self.W_1, self.W_2, self.b_1, self.b_2]]
def sigmoid(self, z):
return 1 / (1 + np.exp(-z))
def ReLU(self, z):
return np.maximum(0, z)
def hidden(self, X, W_1, b):
return self.sigmoid(W_1 @ X.reshape(-1,1) + b)
def output(self, H, W_2, b):
return self.sigmoid(W_2.T @ H + b)
def delta_2(self, o, y):
return float((o - y) * o * (1 - o))
def delta_1(self, W_2, h, d2):
return (W_2 * d2 * h * (1 - h)).reshape(-1, 1)
def dJdW_2(self, h, d2):
return d2 * h
def dJdW_1(self, X, d1):
return d1 @ X.T
def train(self, data, labels, passages=10, lr=0.01):
n=passages//50
for j in range(passages):
#n = int(np.ceil(100*j/passages))
n = round(100*(j+1)/passages)
#print(f'|{n*"█"}{(100-n)*" "}|', end='\r')
print(f'{n*"▣"}{(100-n)*"▢"}', end='\r')
for i in range(data.shape[0]):
h = self.hidden(data[i], self.W_1, self.b_1)
o = self.output(h, self.W_2, self.b_2)
#backward propagation
d2 = self.delta_2(o, labels[i])
djdb_2 = d2
d1 = self.delta_1(self.W_2, h, d2)
djdb_1 = d1
dj2 = self.dJdW_2(h, d2)
dj1 = self.dJdW_1(data[i].reshape(-1, 1), d1)
self.W_2 -= lr * dj2
self.W_1 -= lr * dj1
self.b_2 -= lr * djdb_2
self.b_1 -= lr * djdb_1
self.stock.append([np.copy(self.W_1), np.copy(self.W_2), np.copy(self.b_1), np.copy(self.b_2)])
def predict(self, X, etape=-1):
h = self.hidden(X.reshape(-1, 1), self.stock[etape][0], self.stock[etape][2])
o = self.output(h, self.stock[etape][1], self.stock[etape][3])
return float(o)
def predict_list(self, X, etape=-1):
y_hat = []
for x in X:
y_hat.append(self.predict(x, etape=etape))
return y_hat
def visualisation(self, data, labels, savemod=False, etape=-1, nb_levels=10):
plt.figure(figsize=(12,8))
hh = .02
x_r = data[:, 0].max() - data[:, 0].min()
y_r = data[:, 1].max() - data[:, 1].min()
x_min, x_max = data[:, 0].min() - x_r/10, data[:, 0].max() + x_r/10
y_min, y_max = data[:, 1].min() - y_r/10, data[:, 1].max() + y_r/10
xx, yy = np.meshgrid(np.arange(x_min, x_max, (x_max-x_min)/100),
np.arange(y_min, y_max, (y_max-y_min)/100))
Z = np.array(self.predict_list(np.c_[xx.ravel(), yy.ravel()], etape=etape))
Z = Z.reshape(xx.shape)
contour = plt.contourf(xx, yy, Z, alpha=1, levels=[i/nb_levels for i in range(nb_levels+1)])
sns.scatterplot(x=data[:, 0], y=data[:, 1], hue=labels).set_aspect('equal')
plt.colorbar(contour, label = 'prediction')
if savemod:
plt.savefig(f'./Saves/fig{etape:03d}.png')
else:
plt.show()
def history(self):
return self.stock
